Anthropic’s new research paper puts two numbers side by side. Robots could do 74% of physical tasks. They cost less than the human doing the work on 0.3% of them. One number makes headlines and the other decides if you sign a lease. On robots vs human labour cost, the 0.3% belongs in your budget, and the 74% belongs nowhere near it.
The paper also estimates robots could do tasks making up 34% of US working hours. Same logic. That’s what a robot can do, not what it’s worth doing, and the gap between the two is the entire purchasing decision.
Key points
- Why does robots vs human labour cost beat capability? Capability counts a task as exposed if a robot could perform it at human speed and reliability, usually in a controlled setting.
- Where do robots already win on cost? Almost nowhere, and the best case is narrow. The paper names packers and packagers as the largest cost-competitive occupation.
- What belongs in the all-in annual cost of a robot? More than the hardware. Anthropic’s cost model spreads fixed costs over about 10 years at an 8% cost of capital, and counts integration, maintenance, software, energy, oversight, insurance and decommissioning.
- How would a small firm run the comparison? Take the task, not the job. Start with the task’s share of the person’s pay, find the robot’s all-in annual cost, add the person who still supervises it, and compare.
Why does robots vs human labour cost beat capability?
Capability counts a task as exposed if a robot could perform it at human speed and reliability, usually in a controlled setting. That says nothing about price. Only 1% of US work time is doable by robots in unstructured settings like open roads.
A claim that robots can do three-quarters of physical tasks invites a budget line, and the paper’s own caveats cut against it.
It lists capability limits as blocking about 70% of physical tasks, with manipulation as the biggest gap, and regulation and human preference as further brakes. A task a robot can technically do and a task you can profitably hand it are different lists. Buyers who confuse them end up with an expensive machine and the same headcount.
Where do robots already win on cost?
Almost nowhere, and the best case is narrow. The paper names packers and packagers as the largest cost-competitive occupation. A robot costs about $45,000 a year to replace one worker, against about $49,000 in labour cost, and robots handle 97% of the time-weighted work. Employment in that occupation has fallen 22% since 2015.
Other examples sit on the wrong side of the line. Welders would cost about five times as much to replace with a robot. Cleaners and dishwashers are several times more expensive. Robotaxis are estimated at roughly $7,000 a year more than taxi drivers, with regulation as the real obstacle.
So robots win where the work is repetitive, indoors and standardized, and even then the margin is thin. The packer saving is $4,000 on a $49,000 base, about 8%. Add a software fee or a service call and it’s gone.
What belongs in the all-in annual cost of a robot?
More than the hardware. Anthropic’s cost model spreads fixed costs over about 10 years at an 8% cost of capital, and counts integration, maintenance, software, energy, oversight, insurance and decommissioning. Most quotes lead with the hardware and bury the rest.
Counting the whole cost of ownership is the same discipline as cost per successful task for software. Decommissioning looks like a footnote and isn’t one. When Figure retired its older F.02 humanoids, it said a manual teardown would take its technical staff too long, and the units ended up melted at a foundry in Finland. That is a vendor’s own end-of-life bill, and your lease will have one too.
How would a small firm run the comparison?
Take the task, not the job. Start with the task’s share of the person’s pay, find the robot’s all-in annual cost, add the person who still supervises it, and compare. If the robot isn’t clearly cheaper at the end, the answer is no. The numbers below are ours, for illustration, not the report’s.
Say packing is half of a $60,000 role. The labour cost of the task is $30,000. The vendor quotes $26,000 a year all-in, lease term included. That looks like a $4,000 saving. But someone still supervises and fixes the machine, and if that’s a tenth of another $60,000 role, you’ve added $6,000. The robot now costs $32,000 against $30,000 of labour. You’ve lost $2,000 a year and gained a contract. If public money is paying for the machines, AI funding in Canada mostly comes as a loan, so price the payback first. If you can get help paying for it, Mitacs AI Advantage starts at $5,000, though supervision is the real cost.
| Line | Per year |
|---|---|
| Labour cost of the packing task | $30,000 |
| Vendor quote, all-in | $26,000 |
| Supervision, a tenth of a $60,000 role | $6,000 |
| Robot total | $32,000 |
| Result | $2,000 a year worse |
Illustration from the example above, our arithmetic.
Change one input and the answer flips, which is useful. If the vendor drops to $22,000, the robot costs $28,000 against $30,000 and wins by $2,000. So put one question to the vendor. What number will they sign to, in writing, with oversight and decommissioning inside it.
Most small firms will stop at the supervision line. That’s the method working. The paper itself notes that cost parity doesn’t automatically mean fewer jobs, since supervision and repair work may remain.
What does a falling price mean for the lease you sign?
It means a long contract on a falling price is a bet against yourself. In the report’s model, a 20% fall in robot prices would make robots competitive for about 2.8 million workers. Reaching 10% of human work needs roughly a 70% cost drop.
At 3% a year that’s about 40 years, though the authors’ fast scenario pulls it to 2050. Software prices move on the same logic, as GPT-5.5 pricing per finished task shows.
Second-order effect. If prices fall slowly, as the base case says, you’re not racing anyone and can wait for a quote that clears the bar. If they fall fast, the machine you lease today is overpriced by year three. Either way, a shorter lease or an upgrade clause is worth more than a discount on a long contract. Long commitments on falling-price technology are what off-balance-sheet AI spending tends to hide. Our look at what AI costs a Canadian business makes the same point about software, and the hidden time tax of AI explains why the supervision line keeps surprising owners.
Treat the source with care. Anthropic built its exposure scores using its own Claude model to rate tasks against the O*NET occupation database, and the authors say those ratings rest on Claude’s judgment and on terse task descriptions. Costs are approximate, they skip real-world friction like borrowing limits, and the scenarios apply one price decline to every task. Anthropic also has an obvious interest in the topic, and independent coverage of the paper is thin. Read the figures as one lab’s model until others check them.
The sceptic says averages are useless to a single employer. A firm with high wages, an enclosed workflow and cheap capital could sit far above the 0.3% line, and the report says AI-powered robots might do work that today’s can’t. Both points are right. They don’t change the method. If your own numbers beat the average, the four-line check shows it in an afternoon, and you’ll have a quote to hold the vendor to. We’d drop our rule if a vendor could show an audited all-in cost, oversight included, below your wage bill. Then buy it.
Until then, run the comparison on one task, with your own wage and a real lease quote, before you ask for a demo.
Frequently asked questions
Can robots replace human workers on cost today?
For most tasks, no. Anthropic’s report finds robots are cheaper than the human doing the work on only about 0.3% of tasks, with packers closest to parity.
How much does a robot cost compared with a packer?
The report estimates about $45,000 a year for a robot doing a packer’s work, against about $49,000 in labour cost. That figure includes integration, maintenance, oversight and insurance.
How should a small business price a robot?
Compare the robot’s full annual cost, including the person who supervises it, against the labour cost of the specific task. Choose short leases while prices are falling.
Archive entry dated October 3, 2026, written and fact-checked on October 8, 2026. Sources are linked on the claims they support.
